Using machine learning to advance next-day probabilistic convective hazard prediction with convection-permitting models.

Abstract:

Direct prediction of convective hazards using atmospheric models is difficult due to their small spatial scale. While convection-permitting (CP) numerical weather prediction forecasts can partially resolve convective structures, convective hazards (e.g., tornadoes or hail storms) are typically unresolved with horizontal grid-spacings >= 1 km. Using CP model guidance to provide hazard predictions thus relies on inferring hazard likelihood via surrogate diagnostics, such as updraft speed or updraft helicity (UH). Further, CP models often err in the placement, initiation, and intensity of convective storms, requiring the use of probabilities to convey forecast uncertainty of convective hazards.

Here, we report on efforts to improve probabilistic predictions of hazards by combining surrogate diagnostics and environmental information within a machine learning framework for 1-2 day hazard prediction. Specifically, a feedforward neural network (NN) was used to produce gridded probabilistic convective hazard predictions over the contiguous United States. Input fields to the NN included 174 predictors, derived from 38 environmental and surrogate fields output by ~500 deterministic CP model forecasts, with observed severe storm reports used for training and verification.

To evaluate the skill of the NNs, NN probability forecasts (NNPFs) were compared to surrogate-severe probability forecasts (SSPFs), generated by smoothing a field of surrogate reports derived with UH. NNPFs and SSPFs were produced each forecast hour on an 80-km grid, with forecasts valid for the occurrence of any severe weather report within 40 or 120 km, and 2 h, of each 80-km grid box. In aggregate, NNPFs were superior to SSPFs, producing statistically significant improvements in forecast reliability and resolution. NNPFs were most skillful relative to SSPFs when predicting hazards on larger-scales (e.g., 120 km vs. 40 km) and in situations where using UH was detrimental to forecast skill. These included model spin-up, nocturnal periods, and regions and environments where supercells were less common, such as the western and eastern United States and high-shear, low-CAPE regimes. Comparing ML hazard predictions to a non-ML baseline is an important step toward understanding ML forecast behavior and putting their skill in context.